Extreme Parkour with Legged Robots

RL system for legged robot parkour trained in under 20 hours

License

BSD-3-Clause

Overview

Reinforcement learning system enabling legged robots to perform parkour maneuvers including climbing over obstacles, leaping across gaps, and navigating challenging terrain. Trained in under 20 hours using Isaac Gym simulator.

This groundbreaking research from ICRA 2024 demonstrates how legged robots can learn complex parkour skills through reinforcement learning in remarkably short training times. The system uses a two-stage training pipeline: base policy training (8-10 hours) followed by distillation policy training. Built on Isaac Gym with PyTorch, the framework enables robots to perform nonstop climbing, leaping, and jumping maneuvers that were previously thought to require extensive hand-engineering. The approach combines the legged_gym environment with the rsl_rl reinforcement learning library, supporting both simulator visualization and web-based viewing for headless systems. Models can be exported as JIT-traced versions for deployment on edge devices, making the technology practical for real-world applications.

Why it matters

Extreme Parkour demonstrated that reinforcement learning can teach legged robots to perform agile maneuvers that rival trained animals. The system learned to climb, leap, and crawl through obstacles in under 20 hours of training — a breakthrough in legged robot agility.

Problem it solves

Legged robots are typically limited to flat-ground walking. This project shows that RL can discover agile locomotion behaviors that let robots traverse extreme obstacles, bringing us closer to robots that can navigate any terrain.

Key takeaways

• Legged robots performing parkour: climbing, leaping, crawling through obstacles
• Trained entirely in simulation using reinforcement learning
• Under 20 hours of training to achieve extreme maneuvers
• Successfully transferred to real Unitree robots zero-shot
• No pre-programmed gaits — all behaviors learned end-to-end

Who it is for

• Legged locomotion researchers
• Teams developing agile quadruped controllers
• RL researchers studying complex motor control
• Companies building robots for rough terrain deployment

Limitations

• Requires specific simulation setup for training
• Real-world transfer limited to specific Unitree robot hardware
• Extreme maneuvers may cause hardware wear or damage
• Current approach is task-specific rather than general-purpose